Decomposition in hidden Markov models for activity recognition

Dynamic probabilistic networks have been widely used in activity recognition. However, few models are competent for long-term complex activities involving multi-person interactions. Based on the study of activity characteristics, this paper proposes a decomposed hidden Markov model (DHMM) to capture the structures of activity both in time and space. The model combines spatial decomposition and hierarchical abstraction to reduce the complexity of state space as well as the number of parameters greatly, with consequent computational benefits in efficiency and accuracy. Experiments on two-person interactions and individual activities demonstrate that DHMMs are more powerful than Coupled HMMs and Hierarchical HMMs.